Research And Application Of Face Privacy Protection Algorithm Based On Image Generation Model

In recent years,the rapid development of network media and digital imaging technology has made it easier to obtain and share face image data,especially with the breakthrough of deep learning,face recognition,face payment,monitoring systems and other applications in multiple scenarios The technology is also becoming more and more mature,and various applications related to face images are becoming more and more common in daily life,which leads to the excessive collection of face images in various network applications,which greatly affects the privacy and security of face information.threat.At present,face privacy protection has become a research hotspot in the field of computer vision,which has important research value for strengthening network information security.Aiming at the privacy and security issues of face images in different scenarios,this paper combines the deep learning image generation model to carry out the research and application of face privacy protection algorithms.The main contributions are as follows:(1)For application scenarios such as social sharing and surveillance systems that do not require high levels of identity information,this paper proposes a face anonymization method based on the consistency of facial poses.The method designs a conditional self-encoder,where the input face image is first pre-processed to remove the face facial privacy region,learns the feature information in the original face through a feature extraction network,will pass a small feature mapping network to convert the extracted feature information into a style vector satisfying a normal distribution,and subsequently generates the face image through a Style GAN2 network.Experimental results on several public face datasets show that the proposed model can effectively anonymise face images and protect face privacy information.The proposed model can also achieve good anonymisation results for face images collected in open scenes.(2)For scenarios that require high identity information such as face access control systems and face payment,this paper proposes a face attribute domain transfer method based on identity preservation.In this method,a conditional generative adversarial network is designed.The encoder based on the Vi T structure can better map the image space to the feature space,and then generate the deprived face image through the transposed convolution.Local facial attributes and global facial attributes choose one or more of them to confuse.An identity preservation module is designed in the model,so that the generated face image retains the original identity information while erasing the facial attribute information.The experimental results show that the method can effectively preserve the identity information of the original image,and protect the privacy of the facial attributes of the face,and the image generation effect is more realistic.(3)For different privacy needs in different scenarios,this paper designs and develops a face privacy protection system based on B/S architecture to integrate and apply the face anonymity protection algorithm proposed in this paper.The system mainly consists of three modules: data acquisition,data pre-processing and data generation,which implement the acquisition of user images,face alignment and deprivacy image generation respectively,divided into two image usage scenarios: face anonymisation and identity preservation.It also optimises the data acquisition method,generation speed and robustness of the generation model for open scenarios,which can better meet the practical application needs of users for face privacy protection in multiple scenarios.